Closed Al-Hussein-96 closed 3 years ago
@SilvioGiancola thank you for your response, yes I'm trying to train with labels.json on colab, But the ram is not enough for training Can you tell me the minimum requirements? if i buy colab pro can i train the features on model V1 or V2 ?
You'll need at least 64GB of RAM for training, because the code is loading all the features first.
You'll need at least 64GB of RAM for training, because the code is loading all the features first.
Is there a way to reduce the use of RAM ? because the maximum RAM I have is 30 GB on colab pro
You will have to change the pytorch dataset and dataloader to only read the features when needed (__getitem__()
), that will involve a lot of reading and might slow down the training process.
You will have to change the pytorch dataset and dataloader to only read the features when needed (
__getitem__()
), that will involve a lot of reading and might slow down the training process.
Ok, thanks you very much
Hi @Al-Hussein-96 ,
Note that SoccerNet V1 is deprecated, and we recommend you to focus on the 17 classes of the V2. You can train on the 17 classes of the V2 and use your model only for 3 classes, discarding the results for the other classes. Also, feel free to use the original implementation of CALF on the V1 available here: https://github.com/cioppaanthony/context-aware-loss .
That being said, you should be able to train for the V1 with only a few changes on in this pytorch implementation :
Labels-v2.json
toLabels.json
,model.py
outputs the probability for the 3 classes only instead of the 17. In the code, look for any reference of the number of classes and make sure they only return 3 instead of 17.Good luck with your re-implementation.